Is this a real client?

No. It is an illustrative scenario built from the typical shape of support-copilot projects, to show our approach. Names and figures are examples.

Why not let the AI answer customers directly?

It can be the next step once accuracy on your data is proven and the unanswerable cases are handled well. Starting with agent drafts gives you measured quality before any customer sees an AI answer. See AI support chatbot that doesn’t embarrass you.

Example project · RAG · Customer support

A support copilot for a B2B SaaS: drafts agents actually use.

“Fieldbook” (a fictional company) sells scheduling software to field-service businesses. Its 35 support agents spend much of each ticket searching the help centre and old tickets for the right answer. Here is how we would take a support copilot from idea to daily use — including the parts that usually go wrong.

By Nythrex EngineeringUpdated 24 September 2026 2 min read

Illustrative scenario. This is an illustrative scenario that shows how Nythrex approaches this type of project. It is not a specific client case study: names, figures and timelines are examples.

Starting point

support agents in two time zones

~1,800

help-centre articles, some outdated

3 years

of resolved tickets in Zendesk

product lines with different terminology

Example figures for an illustrative scenario.

Previous attempt: a generic chatbot plugin trained on the help centre. It answered confidently, cited nothing, and agents stopped using it within weeks because they couldn’t tell good answers from bad ones. That history shaped the brief: drafts with sources, for agents, not for customers.

Discovery: two weeks

1
Shadowing agents
Watching real tickets showed that the answer often existed in a resolved ticket, not in the help centre — especially for edge cases and integrations.
2
Test set
Senior agents picked 150 recent tickets across product lines and wrote what a good reply must contain. Ten were deliberately unanswerable from existing content.
3
Data review
Resolved tickets contained customer names, phone numbers and occasionally credentials pasted by customers — a PII-scrubbing step was needed before indexing.
4
Success criteria
Agreed up front: most drafts usable with light edits, correct citation, answer in under 10 seconds, and an honest “no confident answer” on unanswerable tickets.

Proof of concept: four weeks

The first version searched only help articles and did poorly on integration questions. Adding cleaned resolved tickets as a second source — ranked below official articles — changed the picture more than any prompt change. Hybrid search mattered because customers quote exact error codes and plan names that pure vector search misses.

Help articles only, vector search

48%

+ hybrid keyword search

58%

+ resolved tickets as a source

71%

+ reranker & prompt revisions

79%

Illustrative example: share of test-set drafts rated “usable with light edits” by senior agents at each PoC iteration.

How iterations typically move the numbers: retrieval changes dominate.

Architecture

Agent interface

Zendesk sidebar app
Insert draft
Open sources
👍 / 👎 + reason

Copilot service

Ticket summarisation
Query rewriting
Hybrid retrieval
Reranking
Draft with citations

Guardrails

No auto-send
PII redaction
Confidence threshold
Tone & policy checks

Ingestion

Help centre sync
Resolved tickets (scrubbed)
Release notes
Nightly re-index

Client’s Azure tenant (EU)

Azure OpenAI
Azure AI Search
App Service
Key Vault
Log Analytics

Rollout

Discovery

Test set · data review

Proof of concept

Go decision

Production build

Zendesk app · guardrails

Pilot: 6 agents

Daily feedback

All agents

Training · dashboards

0816 weeks

Illustrative plan. Each gate was a decision, not a formality.

What a good outcome looks like

Most

drafts used with light edits

Every

draft links to its sources

Faster

onboarding for new agents

↓

outdated articles, fixed by content owners

Qualitative outcomes of an illustrative scenario — not measured results.

Lessons that transfer to other projects

Agents adopt drafts they can verify. Citations did more for trust than any accuracy number.
Resolved tickets are a goldmine — and a privacy risk. Scrub before you index.
The copilot is also a content audit. Articles it cited but agents rejected were usually outdated.
A future fine-tuning candidate emerged: ticket categorisation, where agents’ corrections are building a labelled dataset for free. See RAG vs fine-tuning.

Frequently asked questions

ServiceRAG developmentRAG development: AI assistants that answer from your documents and data with citations, permissions and measured accuracy — built in your cloud.GuideAn AI support chatbot that doesn’t embarrass youBuild an AI support chatbot customers trust: grounded answers, clear scope, human hand-off, safe actions, abuse protection and honest metrics.GuideRAG vs fine-tuning vs promptingRAG adds knowledge, fine-tuning changes behaviour, prompting is where you start. How to choose — with a decision tree, costs and pitfalls.GuideLLM evals for non-ML teamsHow to evaluate an LLM feature without a data-science team: build a test set, choose metrics, use LLM-as-judge carefully, run evals in CI and read the results.Example projectDocument AI for a distributorIllustrative document AI project: invoices and waybills extracted, matched to purchase orders and posted to the ERP, with human review where it matters.

Want a second opinion on your project?

Tell us what you’re building and where you’re stuck. We’ll reply within one business day with the most practical next step — even if that step isn’t us.

Start a project